*refactoring of SynetConvolution32fNhwcDepthwise class.

ermig1979 · Oct 1, 2024 · f78c4bf · f78c4bf
1 parent 12333d2
commit f78c4bf
Show file tree

Hide file tree

Showing 13 changed files with 460 additions and 2,192 deletions.
diff --git a/docs/2024.html b/docs/2024.html
@@ -39,7 +39,7 @@ <h3 id="R143">November X, 2024 (version X.X.143)</h3>
 <h4>Algorithms</h4>
 <h5>Improving</h5>
 <ul>
- <li>AVX-512BW optimizations of function ConvolutionDirectNhwcConvolutionBiasActivationDepthwise.</li>
+ <li>AVX-512BW optimizations of function Convolution32fNhwcDepthwiseDefault.</li>
 </ul>
 
 <a href="#HOME">Home</a>

diff --git a/src/Simd/SimdAvx2SynetConvolution32f.cpp b/src/Simd/SimdAvx2SynetConvolution32f.cpp
@@ -43,8 +43,8 @@ namespace Simd
                 return new Avx2::SynetConvolution32fDirectNchw(param);
             else if (SynetConvolution32fNhwcDirect::Preferable(param))
                 return new SynetConvolution32fNhwcDirect(param);
-            else if (SynetConvolution32fDirectNhwc::Preferable(param))
-                return new SynetConvolution32fDirectNhwc(param);
+            else if (SynetConvolution32fNhwcDepthwise::Preferable(param))
+                return new SynetConvolution32fNhwcDepthwise(param);
             else if (SynetConvolution32fNhwcGroupedBlock1x2::Preferable(param))
                 return new SynetConvolution32fNhwcGroupedBlock1x2(param);
             else

diff --git a/src/Simd/SimdAvx2SynetConvolution32fNhwcDepthwise.cpp b/src/Simd/SimdAvx2SynetConvolution32fNhwcDepthwise.cpp
diff --git a/src/Simd/SimdAvx512bwSynetConvolution32f.cpp b/src/Simd/SimdAvx512bwSynetConvolution32f.cpp
@@ -43,8 +43,8 @@ namespace Simd
                 return new SynetConvolution32fDirectNchw(param);
             else if (SynetConvolution32fNhwcDirect::Preferable(param))
                 return new SynetConvolution32fNhwcDirect(param);
-            else if (SynetConvolution32fDirectNhwc::Preferable(param))
-                return new SynetConvolution32fDirectNhwc(param);
+            else if (SynetConvolution32fNhwcDepthwise::Preferable(param))
+                return new SynetConvolution32fNhwcDepthwise(param);
             else if (SynetConvolution32fNhwcGroupedBlock1x2::Preferable(param))
                 return new SynetConvolution32fNhwcGroupedBlock1x2(param);
             else

diff --git a/src/Simd/SimdAvx512bwSynetConvolution32fNhwcDepthwise.cpp b/src/Simd/SimdAvx512bwSynetConvolution32fNhwcDepthwise.cpp
diff --git a/src/Simd/SimdBaseSynetConvolution32f.cpp b/src/Simd/SimdBaseSynetConvolution32f.cpp
@@ -46,8 +46,8 @@ namespace Simd
                 return new SynetConvolution32fDirectNchw(param);
             else if (SynetConvolution32fNhwcDirect::Preferable(param))
                 return new SynetConvolution32fNhwcDirect(param);
-            else if (SynetConvolution32fDirectNhwc::Preferable(param))
-                return new SynetConvolution32fDirectNhwc(param);
+            else if (SynetConvolution32fNhwcDepthwise::Preferable(param))
+                return new SynetConvolution32fNhwcDepthwise(param);
             else if (SynetConvolution32fNhwcGroupedBlock1x2::Preferable(param))
                 return new SynetConvolution32fNhwcGroupedBlock1x2(param);
 #endif

diff --git a/src/Simd/SimdBaseSynetConvolution32fNhwcDepthwise.cpp b/src/Simd/SimdBaseSynetConvolution32fNhwcDepthwise.cpp
@@ -32,38 +32,7 @@ namespace Simd
 #if defined(SIMD_SYNET_ENABLE)
     namespace Base
     {
-        SynetConvolution32fDirectNhwc::SynetConvolution32fDirectNhwc(const ConvParam & p)
-            : SynetConvolution32f(p)
-        {
-            _batch = p.batch;
-            _sizeS = p.srcC*p.srcH*p.srcW;
-            _sizeD = p.dstC*p.dstH*p.dstW;
-            _convolutionBiasActivation = SetConvolutionBiasActivation();
-        }
-
-        void SynetConvolution32fDirectNhwc::Forward(const float * src, float * buf, float * dst)
-        {
-            for (size_t b = 0; b < _batch; ++b)
-            {
-                _convolutionBiasActivation(src, _param, _weight, _bias, _params, dst);
-                src += _sizeS;
-                dst += _sizeD;
-            }
-        }
-
-        bool SynetConvolution32fDirectNhwc::Preferable(const ConvParam & p)
-        {
-            if (p.trans == 0)
-                return false;
-            if (p.group == 1)
-            {
-                double k = double(p.srcC) / p.group * p.strideX * p.strideY / p.kernelX / p.kernelY;
-                return k < 2.0;
-            }
-            return p.IsDepthwise();
-        }
-
-        static void ConvolutionDirectNhwcConvolutionBiasActivationDefault(const float * src, const ConvParam & p, const float * weight, const float * bias, const float * params, float * dst)
+        static void ConvolutionNhwcDepthwiseDefault(const float * src, const ConvParam & p, const float * weight, const float * bias, const float * params, float * dst)
         {
             size_t group = p.group;
             size_t srcC = p.srcC / group;
@@ -111,10 +80,31 @@ namespace Simd
             }
         }
 
-        SynetConvolution32fDirectNhwc::ConvolutionBiasActivationPtr SynetConvolution32fDirectNhwc::SetConvolutionBiasActivation()
+        //-------------------------------------------------------------------------------------------------
+
+        SynetConvolution32fNhwcDepthwise::SynetConvolution32fNhwcDepthwise(const ConvParam & p)
+            : SynetConvolution32f(p)
         {
-            return ConvolutionDirectNhwcConvolutionBiasActivationDefault;
+            _batch = p.batch;
+            _sizeS = p.srcC*p.srcH*p.srcW;
+            _sizeD = p.dstC*p.dstH*p.dstW;
+            _convolution = ConvolutionNhwcDepthwiseDefault;
         }
+
+        void SynetConvolution32fNhwcDepthwise::Forward(const float * src, float * buf, float * dst)
+        {
+            for (size_t b = 0; b < _batch; ++b)
+            {
+                _convolution(src, _param, _weight, _bias, _params, dst);
+                src += _sizeS;
+                dst += _sizeD;
+            }
+        }
+
+        bool SynetConvolution32fNhwcDepthwise::Preferable(const ConvParam & p)
+        {
+            return p.trans && p.IsDepthwise();
+        }        
     }
 #endif
 }
diff --git a/src/Simd/SimdNeonSynetConvolution32f.cpp b/src/Simd/SimdNeonSynetConvolution32f.cpp
@@ -50,8 +50,8 @@ namespace Simd
                 return new SynetConvolution32fGemmNT(param);
             else if (SynetConvolution32fNhwcDirect::Preferable(param))
                 return new SynetConvolution32fNhwcDirect(param);
-            else if (SynetConvolution32fDirectNhwc::Preferable(param))
-                return new SynetConvolution32fDirectNhwc(param);
+            else if (SynetConvolution32fNhwcDepthwise::Preferable(param))
+                return new SynetConvolution32fNhwcDepthwise(param);
             else if (Base::SynetConvolution32fNhwcGroupedBlock1x2::Preferable(param))
                 return new Base::SynetConvolution32fNhwcGroupedBlock1x2(param);
             else